Fault diagnosis method for harmonic reducer based on federated prototype domain adaptation with adaptive aggregation strategy

To tackle the poor performance of global model owing to the significant distribution discrepancies in harmonic reducers vibration data among users and limited availability of labeled data under certain operating conditions, a federated prototype domain adaptation framework with an adaptive aggregation strategy for fault diagnosis of harmonic reducers (AFPDA) is proposed. First, the user model parameters and class prototypes are considered to be the core interactive information in the federated learning framework. A multiuser prototype collaboration and interaction mechanism is constructed to enable effective knowledge interaction among users in source domain and target domain. Second, a class-prototype-guided local model training strategy is designed to align local data features with class prototypes while ensuring data privacy. Finally, a divergence-aware adaptive model aggregation strategy is proposed to boost the diagnostic capability of the global model. Extensive comparative experiments are conducted on a bearing dataset and a self-collected harmonic reducer dataset. Compared with other federated transfer learning methods, the AFPDA improves diagnostic accuracy on unlabeled target domain users by at least 3.58 and 7.54%, respectively.